Confidential AI Network

Train on confidential data without moving it into a shared lake

Data providers, model builders, and clean-room operators agree a Ricardian contract, then train only where that contract allows. A Local provider trains on this machine. An Azure or GCP provider submits the job to the cluster they already run.

Open the app Product tour Research notes

Use the cluster you already run

The contract names the cluster. The kubeconfig path is a file on the provider’s host, saved from the TSP dashboard Cloud Provider card. It is not stored on the contract. A missing path fails the job. No subscription, project, or cluster is created for the job.

Local

Training stays on this host. The contract wizard hides the cluster sections.

Azure

Existing AKS cluster: subscription, resource group, cluster name, namespace, region, storage account, and Key Vault URI.

Google Cloud

Existing GKE cluster: project, location, cluster name, namespace, and GCS bucket.

Oracle Cloud

Existing OKE cluster. AWS does not start a training job.

Deploy the trained model where the consumer runs

Training follows the clean-room provider on the contract. After the model is registered, the consumer picks a deploy target: Local, Azure, GCP, or AWS. Local scores on this host. The three clouds submit a Job to a cluster the consumer already named, with a kubeconfig path on the API host. A missing path leaves the model undeployed. The score is not copied back from the cluster, and the run is marked simulated until a hardware quote is attached.

How a job starts

Clean-room screenshots · Azure and GCP deck · Azure confidential computing

Open the product

The app landing page is where you register, sign in, and run a contract. These notes stay here for the architecture and the security write-ups.

Open the app Read the notes